These changes are the raw update to linux-4.4.6-rt14. Kernel sources

are taken from kernel.org, and rt patch from the rt wiki download page.

During the rebasing, the following patch collided:

Force tick interrupt and get rid of softirq magic(I70131fb85).

Collisions have been removed because its logic was found on the
source already.

Change-Id: I7f57a4081d9deaa0d9ccfc41a6c8daccdee3b769
Signed-off-by: José Pekkarinen <jose.pekkarinen@nokia.com>

8 files changed, 1073 insertions, 345 deletions

diff --git a/kernel/arch/arc/mm/Makefile b/kernel/arch/arc/mm/Makefile
index ac95cc239..3703a4969 100644
--- a/kernel/arch/arc/mm/Makefile
+++ b/kernel/arch/arc/mm/Makefile
@@ -7,4 +7,5 @@
 #
 
 obj-y	:= extable.o ioremap.o dma.o fault.o init.o
-obj-y	+= tlb.o tlbex.o cache_arc700.o mmap.o
+obj-y	+= tlb.o tlbex.o cache.o mmap.o
+obj-$(CONFIG_HIGHMEM)	+= highmem.o
diff --git a/kernel/arch/arc/mm/cache_arc700.c b/kernel/arch/arc/mm/cache.c
index 12b2100db..ff7ff6cbb 100644
--- a/kernel/arch/arc/mm/cache_arc700.c
+++ b/kernel/arch/arc/mm/cache.c
@@ -1,64 +1,12 @@
 /*
- * ARC700 VIPT Cache Management
+ * ARC Cache Management
  *
+ * Copyright (C) 2014-15 Synopsys, Inc. (www.synopsys.com)
  * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 as
  * published by the Free Software Foundation.
- *
- *  vineetg: May 2011: for Non-aliasing VIPT D-cache following can be NOPs
- *   -flush_cache_dup_mm (fork)
- *   -likewise for flush_cache_mm (exit/execve)
- *   -likewise for flush_cache_range,flush_cache_page (munmap, exit, COW-break)
- *
- * vineetg: Apr 2011
- *  -Now that MMU can support larger pg sz (16K), the determiniation of
- *   aliasing shd not be based on assumption of 8k pg
- *
- * vineetg: Mar 2011
- *  -optimised version of flush_icache_range( ) for making I/D coherent
- *   when vaddr is available (agnostic of num of aliases)
- *
- * vineetg: Mar 2011
- *  -Added documentation about I-cache aliasing on ARC700 and the way it
- *   was handled up until MMU V2.
- *  -Spotted a three year old bug when killing the 4 aliases, which needs
- *   bottom 2 bits, so we need to do paddr | {0x00, 0x01, 0x02, 0x03}
- *                        instead of paddr | {0x00, 0x01, 0x10, 0x11}
- *   (Rajesh you owe me one now)
- *
- * vineetg: Dec 2010
- *  -Off-by-one error when computing num_of_lines to flush
- *   This broke signal handling with bionic which uses synthetic sigret stub
- *
- * vineetg: Mar 2010
- *  -GCC can't generate ZOL for core cache flush loops.
- *   Conv them into iterations based as opposed to while (start < end) types
- *
- * Vineetg: July 2009
- *  -In I-cache flush routine we used to chk for aliasing for every line INV.
- *   Instead now we setup routines per cache geometry and invoke them
- *   via function pointers.
- *
- * Vineetg: Jan 2009
- *  -Cache Line flush routines used to flush an extra line beyond end addr
- *   because check was while (end >= start) instead of (end > start)
- *     =Some call sites had to work around by doing -1, -4 etc to end param
- *     =Some callers didnt care. This was spec bad in case of INV routines
- *      which would discard valid data (cause of the horrible ext2 bug
- *      in ARC IDE driver)
- *
- * vineetg: June 11th 2008: Fixed flush_icache_range( )
- *  -Since ARC700 caches are not coherent (I$ doesnt snoop D$) both need
- *   to be flushed, which it was not doing.
- *  -load_module( ) passes vmalloc addr (Kernel Virtual Addr) to the API,
- *   however ARC cache maintenance OPs require PHY addr. Thus need to do
- *   vmalloc_to_phy.
- *  -Also added optimisation there, that for range > PAGE SIZE we flush the
- *   entire cache in one shot rather than line by line. For e.g. a module
- *   with Code sz 600k, old code flushed 600k worth of cache (line-by-line),
- *   while cache is only 16 or 32k.
  */
 
 #include <linux/module.h>
@@ -73,9 +21,21 @@
 #include <asm/cachectl.h>
 #include <asm/setup.h>
 
+static int l2_line_sz;
+int ioc_exists;
+volatile int slc_enable = 1, ioc_enable = 1;
+
+void (*_cache_line_loop_ic_fn)(phys_addr_t paddr, unsigned long vaddr,
+			       unsigned long sz, const int cacheop);
+
+void (*__dma_cache_wback_inv)(unsigned long start, unsigned long sz);
+void (*__dma_cache_inv)(unsigned long start, unsigned long sz);
+void (*__dma_cache_wback)(unsigned long start, unsigned long sz);
+
 char *arc_cache_mumbojumbo(int c, char *buf, int len)
 {
 	int n = 0;
+	struct cpuinfo_arc_cache *p;
 
 #define PR_CACHE(p, cfg, str)						\
 	if (!(p)->ver)							\
@@ -86,11 +46,24 @@ char *arc_cache_mumbojumbo(int c, char *buf, int len)
 			(p)->sz_k, (p)->assoc, (p)->line_len,		\
 			(p)->vipt ? "VIPT" : "PIPT",			\
 			(p)->alias ? " aliasing" : "",			\
-			IS_ENABLED(cfg) ? "" : " (not used)");
+			IS_USED_CFG(cfg));
 
 	PR_CACHE(&cpuinfo_arc700[c].icache, CONFIG_ARC_HAS_ICACHE, "I-Cache");
 	PR_CACHE(&cpuinfo_arc700[c].dcache, CONFIG_ARC_HAS_DCACHE, "D-Cache");
 
+	if (!is_isa_arcv2())
+                return buf;
+
+	p = &cpuinfo_arc700[c].slc;
+	if (p->ver)
+		n += scnprintf(buf + n, len - n,
+			       "SLC\t\t: %uK, %uB Line%s\n",
+			       p->sz_k, p->line_len, IS_USED_RUN(slc_enable));
+
+	if (ioc_exists)
+		n += scnprintf(buf + n, len - n, "IOC\t\t:%s\n",
+				IS_DISABLED_RUN(ioc_enable));
+
 	return buf;
 }
 
@@ -99,6 +72,40 @@ char *arc_cache_mumbojumbo(int c, char *buf, int len)
  * the cpuinfo structure for later use.
  * No Validation done here, simply read/convert the BCRs
  */
+static void read_decode_cache_bcr_arcv2(int cpu)
+{
+	struct cpuinfo_arc_cache *p_slc = &cpuinfo_arc700[cpu].slc;
+	struct bcr_generic sbcr;
+
+	struct bcr_slc_cfg {
+#ifdef CONFIG_CPU_BIG_ENDIAN
+		unsigned int pad:24, way:2, lsz:2, sz:4;
+#else
+		unsigned int sz:4, lsz:2, way:2, pad:24;
+#endif
+	} slc_cfg;
+
+	struct bcr_clust_cfg {
+#ifdef CONFIG_CPU_BIG_ENDIAN
+		unsigned int pad:7, c:1, num_entries:8, num_cores:8, ver:8;
+#else
+		unsigned int ver:8, num_cores:8, num_entries:8, c:1, pad:7;
+#endif
+	} cbcr;
+
+	READ_BCR(ARC_REG_SLC_BCR, sbcr);
+	if (sbcr.ver) {
+		READ_BCR(ARC_REG_SLC_CFG, slc_cfg);
+		p_slc->ver = sbcr.ver;
+		p_slc->sz_k = 128 << slc_cfg.sz;
+		l2_line_sz = p_slc->line_len = (slc_cfg.lsz == 0) ? 128 : 64;
+	}
+
+	READ_BCR(ARC_REG_CLUSTER_BCR, cbcr);
+	if (cbcr.c && ioc_enable)
+		ioc_exists = 1;
+}
+
 void read_decode_cache_bcr(void)
 {
 	struct cpuinfo_arc_cache *p_ic, *p_dc;
@@ -117,8 +124,13 @@ void read_decode_cache_bcr(void)
 	if (!ibcr.ver)
 		goto dc_chk;
 
-	BUG_ON(ibcr.config != 3);
-	p_ic->assoc = 2;		/* Fixed to 2w set assoc */
+	if (ibcr.ver <= 3) {
+		BUG_ON(ibcr.config != 3);
+		p_ic->assoc = 2;		/* Fixed to 2w set assoc */
+	} else if (ibcr.ver >= 4) {
+		p_ic->assoc = 1 << ibcr.config;	/* 1,2,4,8 */
+	}
+
 	p_ic->line_len = 8 << ibcr.line_len;
 	p_ic->sz_k = 1 << (ibcr.sz - 1);
 	p_ic->ver = ibcr.ver;
@@ -130,94 +142,135 @@ dc_chk:
 	READ_BCR(ARC_REG_DC_BCR, dbcr);
 
 	if (!dbcr.ver)
-		return;
+		goto slc_chk;
+
+	if (dbcr.ver <= 3) {
+		BUG_ON(dbcr.config != 2);
+		p_dc->assoc = 4;		/* Fixed to 4w set assoc */
+		p_dc->vipt = 1;
+		p_dc->alias = p_dc->sz_k/p_dc->assoc/TO_KB(PAGE_SIZE) > 1;
+	} else if (dbcr.ver >= 4) {
+		p_dc->assoc = 1 << dbcr.config;	/* 1,2,4,8 */
+		p_dc->vipt = 0;
+		p_dc->alias = 0;		/* PIPT so can't VIPT alias */
+	}
 
-	BUG_ON(dbcr.config != 2);
-	p_dc->assoc = 4;		/* Fixed to 4w set assoc */
 	p_dc->line_len = 16 << dbcr.line_len;
 	p_dc->sz_k = 1 << (dbcr.sz - 1);
 	p_dc->ver = dbcr.ver;
-	p_dc->vipt = 1;
-	p_dc->alias = p_dc->sz_k/p_dc->assoc/TO_KB(PAGE_SIZE) > 1;
+
+slc_chk:
+	if (is_isa_arcv2())
+                read_decode_cache_bcr_arcv2(cpu);
 }
 
 /*
- * 1. Validate the Cache Geomtery (compile time config matches hardware)
- * 2. If I-cache suffers from aliasing, setup work arounds (difft flush rtn)
- *    (aliasing D-cache configurations are not supported YET)
- * 3. Enable the Caches, setup default flush mode for D-Cache
- * 3. Calculate the SHMLBA used by user space
+ * Line Operation on {I,D}-Cache
  */
-void arc_cache_init(void)
-{
-	unsigned int __maybe_unused cpu = smp_processor_id();
-	char str[256];
-
-	printk(arc_cache_mumbojumbo(0, str, sizeof(str)));
 
-	if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
-		struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
+#define OP_INV		0x1
+#define OP_FLUSH	0x2
+#define OP_FLUSH_N_INV	0x3
+#define OP_INV_IC	0x4
 
-		if (!ic->ver)
-			panic("cache support enabled but non-existent cache\n");
+/*
+ *		I-Cache Aliasing in ARC700 VIPT caches (MMU v1-v3)
+ *
+ * ARC VIPT I-cache uses vaddr to index into cache and paddr to match the tag.
+ * The orig Cache Management Module "CDU" only required paddr to invalidate a
+ * certain line since it sufficed as index in Non-Aliasing VIPT cache-geometry.
+ * Infact for distinct V1,V2,P: all of {V1-P},{V2-P},{P-P} would end up fetching
+ * the exact same line.
+ *
+ * However for larger Caches (way-size > page-size) - i.e. in Aliasing config,
+ * paddr alone could not be used to correctly index the cache.
+ *
+ * ------------------
+ * MMU v1/v2 (Fixed Page Size 8k)
+ * ------------------
+ * The solution was to provide CDU with these additonal vaddr bits. These
+ * would be bits [x:13], x would depend on cache-geometry, 13 comes from
+ * standard page size of 8k.
+ * H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits
+ * of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the
+ * orig 5 bits of paddr were anyways ignored by CDU line ops, as they
+ * represent the offset within cache-line. The adv of using this "clumsy"
+ * interface for additional info was no new reg was needed in CDU programming
+ * model.
+ *
+ * 17:13 represented the max num of bits passable, actual bits needed were
+ * fewer, based on the num-of-aliases possible.
+ * -for 2 alias possibility, only bit 13 needed (32K cache)
+ * -for 4 alias possibility, bits 14:13 needed (64K cache)
+ *
+ * ------------------
+ * MMU v3
+ * ------------------
+ * This ver of MMU supports variable page sizes (1k-16k): although Linux will
+ * only support 8k (default), 16k and 4k.
+ * However from hardware perspective, smaller page sizes aggrevate aliasing
+ * meaning more vaddr bits needed to disambiguate the cache-line-op ;
+ * the existing scheme of piggybacking won't work for certain configurations.
+ * Two new registers IC_PTAG and DC_PTAG inttoduced.
+ * "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs
+ */
 
-		if (ic->line_len != L1_CACHE_BYTES)
-			panic("ICache line [%d] != kernel Config [%d]",
-			      ic->line_len, L1_CACHE_BYTES);
+static inline
+void __cache_line_loop_v2(phys_addr_t paddr, unsigned long vaddr,
+			  unsigned long sz, const int op)
+{
+	unsigned int aux_cmd;
+	int num_lines;
+	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
 
-		if (ic->ver != CONFIG_ARC_MMU_VER)
-			panic("Cache ver [%d] doesn't match MMU ver [%d]\n",
-			      ic->ver, CONFIG_ARC_MMU_VER);
+	if (op == OP_INV_IC) {
+		aux_cmd = ARC_REG_IC_IVIL;
+	} else {
+		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
+		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
 	}
 
-	if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
-		struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
-		int handled;
-
-		if (!dc->ver)
-			panic("cache support enabled but non-existent cache\n");
+	/* Ensure we properly floor/ceil the non-line aligned/sized requests
+	 * and have @paddr - aligned to cache line and integral @num_lines.
+	 * This however can be avoided for page sized since:
+	 *  -@paddr will be cache-line aligned already (being page aligned)
+	 *  -@sz will be integral multiple of line size (being page sized).
+	 */
+	if (!full_page) {
+		sz += paddr & ~CACHE_LINE_MASK;
+		paddr &= CACHE_LINE_MASK;
+		vaddr &= CACHE_LINE_MASK;
+	}
 
-		if (dc->line_len != L1_CACHE_BYTES)
-			panic("DCache line [%d] != kernel Config [%d]",
-			      dc->line_len, L1_CACHE_BYTES);
+	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
 
-		/* check for D-Cache aliasing */
-		handled = IS_ENABLED(CONFIG_ARC_CACHE_VIPT_ALIASING);
+	/* MMUv2 and before: paddr contains stuffed vaddrs bits */
+	paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
 
-		if (dc->alias && !handled)
-			panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
-		else if (!dc->alias && handled)
-			panic("Don't need CONFIG_ARC_CACHE_VIPT_ALIASING\n");
+	while (num_lines-- > 0) {
+		write_aux_reg(aux_cmd, paddr);
+		paddr += L1_CACHE_BYTES;
 	}
 }
 
-#define OP_INV		0x1
-#define OP_FLUSH	0x2
-#define OP_FLUSH_N_INV	0x3
-#define OP_INV_IC	0x4
-
 /*
- * Common Helper for Line Operations on {I,D}-Cache
+ * For ARC700 MMUv3 I-cache and D-cache flushes
+ * Also reused for HS38 aliasing I-cache configuration
  */
-static inline void __cache_line_loop(unsigned long paddr, unsigned long vaddr,
-				     unsigned long sz, const int cacheop)
+static inline
+void __cache_line_loop_v3(phys_addr_t paddr, unsigned long vaddr,
+			  unsigned long sz, const int op)
 {
 	unsigned int aux_cmd, aux_tag;
 	int num_lines;
-	const int full_page_op = __builtin_constant_p(sz) && sz == PAGE_SIZE;
+	const int full_page = __builtin_constant_p(sz) && sz == PAGE_SIZE;
 
-	if (cacheop == OP_INV_IC) {
+	if (op == OP_INV_IC) {
 		aux_cmd = ARC_REG_IC_IVIL;
-#if (CONFIG_ARC_MMU_VER > 2)
 		aux_tag = ARC_REG_IC_PTAG;
-#endif
-	}
-	else {
-		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
-		aux_cmd = cacheop & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
-#if (CONFIG_ARC_MMU_VER > 2)
+	} else {
+		aux_cmd = op & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
 		aux_tag = ARC_REG_DC_PTAG;
-#endif
 	}
 
 	/* Ensure we properly floor/ceil the non-line aligned/sized requests
@@ -226,177 +279,201 @@ static inline void __cache_line_loop(unsigned long paddr, unsigned long vaddr,
 	 *  -@paddr will be cache-line aligned already (being page aligned)
 	 *  -@sz will be integral multiple of line size (being page sized).
 	 */
-	if (!full_page_op) {
+	if (!full_page) {
 		sz += paddr & ~CACHE_LINE_MASK;
 		paddr &= CACHE_LINE_MASK;
 		vaddr &= CACHE_LINE_MASK;
 	}
-
 	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
 
-#if (CONFIG_ARC_MMU_VER <= 2)
-	/* MMUv2 and before: paddr contains stuffed vaddrs bits */
-	paddr |= (vaddr >> PAGE_SHIFT) & 0x1F;
-#else
-	/* if V-P const for loop, PTAG can be written once outside loop */
-	if (full_page_op)
+	/*
+	 * MMUv3, cache ops require paddr in PTAG reg
+	 * if V-P const for loop, PTAG can be written once outside loop
+	 */
+	if (full_page)
 		write_aux_reg(aux_tag, paddr);
-#endif
+
+	/*
+	 * This is technically for MMU v4, using the MMU v3 programming model
+	 * Special work for HS38 aliasing I-cache configuratino with PAE40
+	 *   - upper 8 bits of paddr need to be written into PTAG_HI
+	 *   - (and needs to be written before the lower 32 bits)
+	 * Note that PTAG_HI is hoisted outside the line loop
+	 */
+	if (is_pae40_enabled() && op == OP_INV_IC)
+		write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
 
 	while (num_lines-- > 0) {
-#if (CONFIG_ARC_MMU_VER > 2)
-		/* MMUv3, cache ops require paddr seperately */
-		if (!full_page_op) {
+		if (!full_page) {
 			write_aux_reg(aux_tag, paddr);
 			paddr += L1_CACHE_BYTES;
 		}
 
 		write_aux_reg(aux_cmd, vaddr);
 		vaddr += L1_CACHE_BYTES;
-#else
+	}
+}
+
+/*
+ * In HS38x (MMU v4), I-cache is VIPT (can alias), D-cache is PIPT
+ * Here's how cache ops are implemented
+ *
+ *  - D-cache: only paddr needed (in DC_IVDL/DC_FLDL)
+ *  - I-cache Non Aliasing: Despite VIPT, only paddr needed (in IC_IVIL)
+ *  - I-cache Aliasing: Both vaddr and paddr needed (in IC_IVIL, IC_PTAG
+ *    respectively, similar to MMU v3 programming model, hence
+ *    __cache_line_loop_v3() is used)
+ *
+ * If PAE40 is enabled, independent of aliasing considerations, the higher bits
+ * needs to be written into PTAG_HI
+ */
+static inline
+void __cache_line_loop_v4(phys_addr_t paddr, unsigned long vaddr,
+			  unsigned long sz, const int cacheop)
+{
+	unsigned int aux_cmd;
+	int num_lines;
+	const int full_page_op = __builtin_constant_p(sz) && sz == PAGE_SIZE;
+
+	if (cacheop == OP_INV_IC) {
+		aux_cmd = ARC_REG_IC_IVIL;
+	} else {
+		/* d$ cmd: INV (discard or wback-n-discard) OR FLUSH (wback) */
+		aux_cmd = cacheop & OP_INV ? ARC_REG_DC_IVDL : ARC_REG_DC_FLDL;
+	}
+
+	/* Ensure we properly floor/ceil the non-line aligned/sized requests
+	 * and have @paddr - aligned to cache line and integral @num_lines.
+	 * This however can be avoided for page sized since:
+	 *  -@paddr will be cache-line aligned already (being page aligned)
+	 *  -@sz will be integral multiple of line size (being page sized).
+	 */
+	if (!full_page_op) {
+		sz += paddr & ~CACHE_LINE_MASK;
+		paddr &= CACHE_LINE_MASK;
+	}
+
+	num_lines = DIV_ROUND_UP(sz, L1_CACHE_BYTES);
+
+	/*
+	 * For HS38 PAE40 configuration
+	 *   - upper 8 bits of paddr need to be written into PTAG_HI
+	 *   - (and needs to be written before the lower 32 bits)
+	 */
+	if (is_pae40_enabled()) {
+		if (cacheop == OP_INV_IC)
+			/*
+			 * Non aliasing I-cache in HS38,
+			 * aliasing I-cache handled in __cache_line_loop_v3()
+			 */
+			write_aux_reg(ARC_REG_IC_PTAG_HI, (u64)paddr >> 32);
+		else
+			write_aux_reg(ARC_REG_DC_PTAG_HI, (u64)paddr >> 32);
+	}
+
+	while (num_lines-- > 0) {
 		write_aux_reg(aux_cmd, paddr);
 		paddr += L1_CACHE_BYTES;
-#endif
 	}
 }
 
+#if (CONFIG_ARC_MMU_VER < 3)
+#define __cache_line_loop	__cache_line_loop_v2
+#elif (CONFIG_ARC_MMU_VER == 3)
+#define __cache_line_loop	__cache_line_loop_v3
+#elif (CONFIG_ARC_MMU_VER > 3)
+#define __cache_line_loop	__cache_line_loop_v4
+#endif
+
 #ifdef CONFIG_ARC_HAS_DCACHE
 
 /***************************************************************
  * Machine specific helpers for Entire D-Cache or Per Line ops
  */
 
-static inline unsigned int __before_dc_op(const int op)
+static inline void __before_dc_op(const int op)
 {
-	unsigned int reg = reg;
-
 	if (op == OP_FLUSH_N_INV) {
 		/* Dcache provides 2 cmd: FLUSH or INV
 		 * INV inturn has sub-modes: DISCARD or FLUSH-BEFORE
 		 * flush-n-inv is achieved by INV cmd but with IM=1
 		 * So toggle INV sub-mode depending on op request and default
 		 */
-		reg = read_aux_reg(ARC_REG_DC_CTRL);
-		write_aux_reg(ARC_REG_DC_CTRL, reg | DC_CTRL_INV_MODE_FLUSH)
-			;
+		const unsigned int ctl = ARC_REG_DC_CTRL;
+		write_aux_reg(ctl, read_aux_reg(ctl) | DC_CTRL_INV_MODE_FLUSH);
 	}
-
-	return reg;
 }
 
-static inline void __after_dc_op(const int op, unsigned int reg)
+static inline void __after_dc_op(const int op)
 {
-	if (op & OP_FLUSH)	/* flush / flush-n-inv both wait */
-		while (read_aux_reg(ARC_REG_DC_CTRL) & DC_CTRL_FLUSH_STATUS);
+	if (op & OP_FLUSH) {
+		const unsigned int ctl = ARC_REG_DC_CTRL;
+		unsigned int reg;
+
+		/* flush / flush-n-inv both wait */
+		while ((reg = read_aux_reg(ctl)) & DC_CTRL_FLUSH_STATUS)
+			;
 
-	/* Switch back to default Invalidate mode */
-	if (op == OP_FLUSH_N_INV)
-		write_aux_reg(ARC_REG_DC_CTRL, reg & ~DC_CTRL_INV_MODE_FLUSH);
+		/* Switch back to default Invalidate mode */
+		if (op == OP_FLUSH_N_INV)
+			write_aux_reg(ctl, reg & ~DC_CTRL_INV_MODE_FLUSH);
+	}
 }
 
 /*
  * Operation on Entire D-Cache
- * @cacheop = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
+ * @op = {OP_INV, OP_FLUSH, OP_FLUSH_N_INV}
  * Note that constant propagation ensures all the checks are gone
  * in generated code
  */
-static inline void __dc_entire_op(const int cacheop)
+static inline void __dc_entire_op(const int op)
 {
-	unsigned int ctrl_reg;
 	int aux;
 
-	ctrl_reg = __before_dc_op(cacheop);
+	__before_dc_op(op);
 
-	if (cacheop & OP_INV)	/* Inv or flush-n-inv use same cmd reg */
+	if (op & OP_INV)	/* Inv or flush-n-inv use same cmd reg */
 		aux = ARC_REG_DC_IVDC;
 	else
 		aux = ARC_REG_DC_FLSH;
 
 	write_aux_reg(aux, 0x1);
 
-	__after_dc_op(cacheop, ctrl_reg);
+	__after_dc_op(op);
 }
 
 /* For kernel mappings cache operation: index is same as paddr */
 #define __dc_line_op_k(p, sz, op)	__dc_line_op(p, p, sz, op)
 
 /*
- * D-Cache : Per Line INV (discard or wback+discard) or FLUSH (wback)
+ * D-Cache Line ops: Per Line INV (discard or wback+discard) or FLUSH (wback)
  */
-static inline void __dc_line_op(unsigned long paddr, unsigned long vaddr,
-				unsigned long sz, const int cacheop)
+static inline void __dc_line_op(phys_addr_t paddr, unsigned long vaddr,
+				unsigned long sz, const int op)
 {
 	unsigned long flags;
-	unsigned int ctrl_reg;
 
 	local_irq_save(flags);
 
-	ctrl_reg = __before_dc_op(cacheop);
+	__before_dc_op(op);
 
-	__cache_line_loop(paddr, vaddr, sz, cacheop);
+	__cache_line_loop(paddr, vaddr, sz, op);
 
-	__after_dc_op(cacheop, ctrl_reg);
+	__after_dc_op(op);
 
 	local_irq_restore(flags);
 }
 
 #else
 
-#define __dc_entire_op(cacheop)
-#define __dc_line_op(paddr, vaddr, sz, cacheop)
-#define __dc_line_op_k(paddr, sz, cacheop)
+#define __dc_entire_op(op)
+#define __dc_line_op(paddr, vaddr, sz, op)
+#define __dc_line_op_k(paddr, sz, op)
 
 #endif /* CONFIG_ARC_HAS_DCACHE */
 
-
 #ifdef CONFIG_ARC_HAS_ICACHE
 
-/*
- *		I-Cache Aliasing in ARC700 VIPT caches
- *
- * ARC VIPT I-cache uses vaddr to index into cache and paddr to match the tag.
- * The orig Cache Management Module "CDU" only required paddr to invalidate a
- * certain line since it sufficed as index in Non-Aliasing VIPT cache-geometry.
- * Infact for distinct V1,V2,P: all of {V1-P},{V2-P},{P-P} would end up fetching
- * the exact same line.
- *
- * However for larger Caches (way-size > page-size) - i.e. in Aliasing config,
- * paddr alone could not be used to correctly index the cache.
- *
- * ------------------
- * MMU v1/v2 (Fixed Page Size 8k)
- * ------------------
- * The solution was to provide CDU with these additonal vaddr bits. These
- * would be bits [x:13], x would depend on cache-geometry, 13 comes from
- * standard page size of 8k.
- * H/w folks chose [17:13] to be a future safe range, and moreso these 5 bits
- * of vaddr could easily be "stuffed" in the paddr as bits [4:0] since the
- * orig 5 bits of paddr were anyways ignored by CDU line ops, as they
- * represent the offset within cache-line. The adv of using this "clumsy"
- * interface for additional info was no new reg was needed in CDU programming
- * model.
- *
- * 17:13 represented the max num of bits passable, actual bits needed were
- * fewer, based on the num-of-aliases possible.
- * -for 2 alias possibility, only bit 13 needed (32K cache)
- * -for 4 alias possibility, bits 14:13 needed (64K cache)
- *
- * ------------------
- * MMU v3
- * ------------------
- * This ver of MMU supports variable page sizes (1k-16k): although Linux will
- * only support 8k (default), 16k and 4k.
- * However from hardware perspective, smaller page sizes aggrevate aliasing
- * meaning more vaddr bits needed to disambiguate the cache-line-op ;
- * the existing scheme of piggybacking won't work for certain configurations.
- * Two new registers IC_PTAG and DC_PTAG inttoduced.
- * "tag" bits are provided in PTAG, index bits in existing IVIL/IVDL/FLDL regs
- */
-
-/***********************************************************
- * Machine specific helper for per line I-Cache invalidate.
- */
-
 static inline void __ic_entire_inv(void)
 {
 	write_aux_reg(ARC_REG_IC_IVIC, 1);
@@ -404,13 +481,13 @@ static inline void __ic_entire_inv(void)
 }
 
 static inline void
-__ic_line_inv_vaddr_local(unsigned long paddr, unsigned long vaddr,
+__ic_line_inv_vaddr_local(phys_addr_t paddr, unsigned long vaddr,
 			  unsigned long sz)
 {
 	unsigned long flags;
 
 	local_irq_save(flags);
-	__cache_line_loop(paddr, vaddr, sz, OP_INV_IC);
+	(*_cache_line_loop_ic_fn)(paddr, vaddr, sz, OP_INV_IC);
 	local_irq_restore(flags);
 }
 
@@ -421,7 +498,7 @@ __ic_line_inv_vaddr_local(unsigned long paddr, unsigned long vaddr,
 #else
 
 struct ic_inv_args {
-	unsigned long paddr, vaddr;
+	phys_addr_t paddr, vaddr;
 	int sz;
 };
 
@@ -432,7 +509,7 @@ static void __ic_line_inv_vaddr_helper(void *info)
         __ic_line_inv_vaddr_local(ic_inv->paddr, ic_inv->vaddr, ic_inv->sz);
 }
 
-static void __ic_line_inv_vaddr(unsigned long paddr, unsigned long vaddr,
+static void __ic_line_inv_vaddr(phys_addr_t paddr, unsigned long vaddr,
 				unsigned long sz)
 {
 	struct ic_inv_args ic_inv = {
@@ -453,6 +530,56 @@ static void __ic_line_inv_vaddr(unsigned long paddr, unsigned long vaddr,
 
 #endif /* CONFIG_ARC_HAS_ICACHE */
 
+noinline void slc_op(phys_addr_t paddr, unsigned long sz, const int op)
+{
+#ifdef CONFIG_ISA_ARCV2
+	/*
+	 * SLC is shared between all cores and concurrent aux operations from
+	 * multiple cores need to be serialized using a spinlock
+	 * A concurrent operation can be silently ignored and/or the old/new
+	 * operation can remain incomplete forever (lockup in SLC_CTRL_BUSY loop
+	 * below)
+	 */
+	static DEFINE_SPINLOCK(lock);
+	unsigned long flags;
+	unsigned int ctrl;
+
+	spin_lock_irqsave(&lock, flags);
+
+	/*
+	 * The Region Flush operation is specified by CTRL.RGN_OP[11..9]
+	 *  - b'000 (default) is Flush,
+	 *  - b'001 is Invalidate if CTRL.IM == 0
+	 *  - b'001 is Flush-n-Invalidate if CTRL.IM == 1
+	 */
+	ctrl = read_aux_reg(ARC_REG_SLC_CTRL);
+
+	/* Don't rely on default value of IM bit */
+	if (!(op & OP_FLUSH))		/* i.e. OP_INV */
+		ctrl &= ~SLC_CTRL_IM;	/* clear IM: Disable flush before Inv */
+	else
+		ctrl |= SLC_CTRL_IM;
+
+	if (op & OP_INV)
+		ctrl |= SLC_CTRL_RGN_OP_INV;	/* Inv or flush-n-inv */
+	else
+		ctrl &= ~SLC_CTRL_RGN_OP_INV;
+
+	write_aux_reg(ARC_REG_SLC_CTRL, ctrl);
+
+	/*
+	 * Lower bits are ignored, no need to clip
+	 * END needs to be setup before START (latter triggers the operation)
+	 * END can't be same as START, so add (l2_line_sz - 1) to sz
+	 */
+	write_aux_reg(ARC_REG_SLC_RGN_END, (paddr + sz + l2_line_sz - 1));
+	write_aux_reg(ARC_REG_SLC_RGN_START, paddr);
+
+	while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
+
+	spin_unlock_irqrestore(&lock, flags);
+#endif
+}
 
 /***********************************************************
  * Exported APIs
@@ -493,7 +620,7 @@ void flush_dcache_page(struct page *page)
 	} else if (page_mapped(page)) {
 
 		/* kernel reading from page with U-mapping */
-		void *paddr = page_address(page);
+		phys_addr_t paddr = (unsigned long)page_address(page);
 		unsigned long vaddr = page->index << PAGE_CACHE_SHIFT;
 
 		if (addr_not_cache_congruent(paddr, vaddr))
@@ -502,22 +629,74 @@ void flush_dcache_page(struct page *page)
 }
 EXPORT_SYMBOL(flush_dcache_page);
 
+/*
+ * DMA ops for systems with L1 cache only
+ * Make memory coherent with L1 cache by flushing/invalidating L1 lines
+ */
+static void __dma_cache_wback_inv_l1(unsigned long start, unsigned long sz)
+{
+	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
+}
 
-void dma_cache_wback_inv(unsigned long start, unsigned long sz)
+static void __dma_cache_inv_l1(unsigned long start, unsigned long sz)
+{
+	__dc_line_op_k(start, sz, OP_INV);
+}
+
+static void __dma_cache_wback_l1(unsigned long start, unsigned long sz)
+{
+	__dc_line_op_k(start, sz, OP_FLUSH);
+}
+
+/*
+ * DMA ops for systems with both L1 and L2 caches, but without IOC
+ * Both L1 and L2 lines need to be explicity flushed/invalidated
+ */
+static void __dma_cache_wback_inv_slc(unsigned long start, unsigned long sz)
 {
 	__dc_line_op_k(start, sz, OP_FLUSH_N_INV);
+	slc_op(start, sz, OP_FLUSH_N_INV);
+}
+
+static void __dma_cache_inv_slc(unsigned long start, unsigned long sz)
+{
+	__dc_line_op_k(start, sz, OP_INV);
+	slc_op(start, sz, OP_INV);
+}
+
+static void __dma_cache_wback_slc(unsigned long start, unsigned long sz)
+{
+	__dc_line_op_k(start, sz, OP_FLUSH);
+	slc_op(start, sz, OP_FLUSH);
+}
+
+/*
+ * DMA ops for systems with IOC
+ * IOC hardware snoops all DMA traffic keeping the caches consistent with
+ * memory - eliding need for any explicit cache maintenance of DMA buffers
+ */
+static void __dma_cache_wback_inv_ioc(unsigned long start, unsigned long sz) {}
+static void __dma_cache_inv_ioc(unsigned long start, unsigned long sz) {}
+static void __dma_cache_wback_ioc(unsigned long start, unsigned long sz) {}
+
+/*
+ * Exported DMA API
+ */
+void dma_cache_wback_inv(unsigned long start, unsigned long sz)
+{
+	__dma_cache_wback_inv(start, sz);
 }
 EXPORT_SYMBOL(dma_cache_wback_inv);
 
 void dma_cache_inv(unsigned long start, unsigned long sz)
 {
-	__dc_line_op_k(start, sz, OP_INV);
+	__dma_cache_inv(start, sz);
 }
 EXPORT_SYMBOL(dma_cache_inv);
 
 void dma_cache_wback(unsigned long start, unsigned long sz)
 {
-	__dc_line_op_k(start, sz, OP_FLUSH);
+	__dma_cache_wback(start, sz);
 }
 EXPORT_SYMBOL(dma_cache_wback);
 
@@ -589,14 +768,14 @@ EXPORT_SYMBOL(flush_icache_range);
  *    builtin kernel page will not have any virtual mappings.
  *    kprobe on loadable module will be kernel vaddr.
  */
-void __sync_icache_dcache(unsigned long paddr, unsigned long vaddr, int len)
+void __sync_icache_dcache(phys_addr_t paddr, unsigned long vaddr, int len)
 {
 	__dc_line_op(paddr, vaddr, len, OP_FLUSH_N_INV);
 	__ic_line_inv_vaddr(paddr, vaddr, len);
 }
 
 /* wrapper to compile time eliminate alignment checks in flush loop */
-void __inv_icache_page(unsigned long paddr, unsigned long vaddr)
+void __inv_icache_page(phys_addr_t paddr, unsigned long vaddr)
 {
 	__ic_line_inv_vaddr(paddr, vaddr, PAGE_SIZE);
 }
@@ -605,7 +784,7 @@ void __inv_icache_page(unsigned long paddr, unsigned long vaddr)
  * wrapper to clearout kernel or userspace mappings of a page
  * For kernel mappings @vaddr == @paddr
  */
-void ___flush_dcache_page(unsigned long paddr, unsigned long vaddr)
+void __flush_dcache_page(phys_addr_t paddr, unsigned long vaddr)
 {
 	__dc_line_op(paddr, vaddr & PAGE_MASK, PAGE_SIZE, OP_FLUSH_N_INV);
 }
@@ -637,7 +816,7 @@ void flush_cache_page(struct vm_area_struct *vma, unsigned long u_vaddr,
 
 	u_vaddr &= PAGE_MASK;
 
-	___flush_dcache_page(paddr, u_vaddr);
+	__flush_dcache_page(paddr, u_vaddr);
 
 	if (vma->vm_flags & VM_EXEC)
 		__inv_icache_page(paddr, u_vaddr);
@@ -663,8 +842,8 @@ void flush_anon_page(struct vm_area_struct *vma, struct page *page,
 void copy_user_highpage(struct page *to, struct page *from,
 	unsigned long u_vaddr, struct vm_area_struct *vma)
 {
-	void *kfrom = page_address(from);
-	void *kto = page_address(to);
+	void *kfrom = kmap_atomic(from);
+	void *kto = kmap_atomic(to);
 	int clean_src_k_mappings = 0;
 
 	/*
@@ -674,9 +853,12 @@ void copy_user_highpage(struct page *to, struct page *from,
 	 *
 	 * Note that while @u_vaddr refers to DST page's userspace vaddr, it is
 	 * equally valid for SRC page as well
+	 *
+	 * For !VIPT cache, all of this gets compiled out as
+	 * addr_not_cache_congruent() is 0
 	 */
 	if (page_mapped(from) && addr_not_cache_congruent(kfrom, u_vaddr)) {
-		__flush_dcache_page(kfrom, u_vaddr);
+		__flush_dcache_page((unsigned long)kfrom, u_vaddr);
 		clean_src_k_mappings = 1;
 	}
 
@@ -697,11 +879,14 @@ void copy_user_highpage(struct page *to, struct page *from,
 	 * sync the kernel mapping back to physical page
 	 */
 	if (clean_src_k_mappings) {
-		__flush_dcache_page(kfrom, kfrom);
+		__flush_dcache_page((unsigned long)kfrom, (unsigned long)kfrom);
 		set_bit(PG_dc_clean, &from->flags);
 	} else {
 		clear_bit(PG_dc_clean, &from->flags);
 	}
+
+	kunmap_atomic(kto);
+	kunmap_atomic(kfrom);
 }
 
 void clear_user_page(void *to, unsigned long u_vaddr, struct page *page)
@@ -721,3 +906,93 @@ SYSCALL_DEFINE3(cacheflush, uint32_t, start, uint32_t, sz, uint32_t, flags)
 	flush_cache_all();
 	return 0;
 }
+
+void arc_cache_init(void)
+{
+	unsigned int __maybe_unused cpu = smp_processor_id();
+	char str[256];
+
+	printk(arc_cache_mumbojumbo(0, str, sizeof(str)));
+
+	if (IS_ENABLED(CONFIG_ARC_HAS_ICACHE)) {
+		struct cpuinfo_arc_cache *ic = &cpuinfo_arc700[cpu].icache;
+
+		if (!ic->ver)
+			panic("cache support enabled but non-existent cache\n");
+
+		if (ic->line_len != L1_CACHE_BYTES)
+			panic("ICache line [%d] != kernel Config [%d]",
+			      ic->line_len, L1_CACHE_BYTES);
+
+		if (ic->ver != CONFIG_ARC_MMU_VER)
+			panic("Cache ver [%d] doesn't match MMU ver [%d]\n",
+			      ic->ver, CONFIG_ARC_MMU_VER);
+
+		/*
+		 * In MMU v4 (HS38x) the alising icache config uses IVIL/PTAG
+		 * pair to provide vaddr/paddr respectively, just as in MMU v3
+		 */
+		if (is_isa_arcv2() && ic->alias)
+			_cache_line_loop_ic_fn = __cache_line_loop_v3;
+		else
+			_cache_line_loop_ic_fn = __cache_line_loop;
+	}
+
+	if (IS_ENABLED(CONFIG_ARC_HAS_DCACHE)) {
+		struct cpuinfo_arc_cache *dc = &cpuinfo_arc700[cpu].dcache;
+
+		if (!dc->ver)
+			panic("cache support enabled but non-existent cache\n");
+
+		if (dc->line_len != L1_CACHE_BYTES)
+			panic("DCache line [%d] != kernel Config [%d]",
+			      dc->line_len, L1_CACHE_BYTES);
+
+		/* check for D-Cache aliasing on ARCompact: ARCv2 has PIPT */
+		if (is_isa_arcompact()) {
+			int handled = IS_ENABLED(CONFIG_ARC_CACHE_VIPT_ALIASING);
+
+			if (dc->alias && !handled)
+				panic("Enable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
+			else if (!dc->alias && handled)
+				panic("Disable CONFIG_ARC_CACHE_VIPT_ALIASING\n");
+		}
+	}
+
+	if (is_isa_arcv2() && l2_line_sz && !slc_enable) {
+
+		/* IM set : flush before invalidate */
+		write_aux_reg(ARC_REG_SLC_CTRL,
+			read_aux_reg(ARC_REG_SLC_CTRL) | SLC_CTRL_IM);
+
+		write_aux_reg(ARC_REG_SLC_INVALIDATE, 1);
+
+		/* Important to wait for flush to complete */
+		while (read_aux_reg(ARC_REG_SLC_CTRL) & SLC_CTRL_BUSY);
+		write_aux_reg(ARC_REG_SLC_CTRL,
+			read_aux_reg(ARC_REG_SLC_CTRL) | SLC_CTRL_DISABLE);
+	}
+
+	if (is_isa_arcv2() && ioc_exists) {
+		/* IO coherency base - 0x8z */
+		write_aux_reg(ARC_REG_IO_COH_AP0_BASE, 0x80000);
+		/* IO coherency aperture size - 512Mb: 0x8z-0xAz */
+		write_aux_reg(ARC_REG_IO_COH_AP0_SIZE, 0x11);
+		/* Enable partial writes */
+		write_aux_reg(ARC_REG_IO_COH_PARTIAL, 1);
+		/* Enable IO coherency */
+		write_aux_reg(ARC_REG_IO_COH_ENABLE, 1);
+
+		__dma_cache_wback_inv = __dma_cache_wback_inv_ioc;
+		__dma_cache_inv = __dma_cache_inv_ioc;
+		__dma_cache_wback = __dma_cache_wback_ioc;
+	} else if (is_isa_arcv2() && l2_line_sz && slc_enable) {
+		__dma_cache_wback_inv = __dma_cache_wback_inv_slc;
+		__dma_cache_inv = __dma_cache_inv_slc;
+		__dma_cache_wback = __dma_cache_wback_slc;
+	} else {
+		__dma_cache_wback_inv = __dma_cache_wback_inv_l1;
+		__dma_cache_inv = __dma_cache_inv_l1;
+		__dma_cache_wback = __dma_cache_wback_l1;
+	}
+}
diff --git a/kernel/arch/arc/mm/dma.c b/kernel/arch/arc/mm/dma.c
index 12cc6485b..29a46bb19 100644
--- a/kernel/arch/arc/mm/dma.c
+++ b/kernel/arch/arc/mm/dma.c
@@ -14,13 +14,12 @@
  * Cache bit off in the TLB entry.
  *
  * The default DMA address == Phy address which is 0x8000_0000 based.
- * A platform/device can make it zero based, by over-riding
- * plat_{dma,kernel}_addr_to_{kernel,dma}
  */
 
 #include <linux/dma-mapping.h>
 #include <linux/dma-debug.h>
 #include <linux/export.h>
+#include <asm/cache.h>
 #include <asm/cacheflush.h>
 
 /*
@@ -37,7 +36,7 @@ void *dma_alloc_noncoherent(struct device *dev, size_t size,
 		return NULL;
 
 	/* This is bus address, platform dependent */
-	*dma_handle = plat_kernel_addr_to_dma(dev, paddr);
+	*dma_handle = (dma_addr_t)paddr;
 
 	return paddr;
 }
@@ -46,8 +45,7 @@ EXPORT_SYMBOL(dma_alloc_noncoherent);
 void dma_free_noncoherent(struct device *dev, size_t size, void *vaddr,
 			  dma_addr_t dma_handle)
 {
-	free_pages_exact((void *)plat_dma_addr_to_kernel(dev, dma_handle),
-			 size);
+	free_pages_exact((void *)dma_handle, size);
 }
 EXPORT_SYMBOL(dma_free_noncoherent);
 
@@ -56,6 +54,20 @@ void *dma_alloc_coherent(struct device *dev, size_t size,
 {
 	void *paddr, *kvaddr;
 
+	/*
+	 * IOC relies on all data (even coherent DMA data) being in cache
+	 * Thus allocate normal cached memory
+	 *
+	 * The gains with IOC are two pronged:
+	 *   -For streaming data, elides needs for cache maintenance, saving
+	 *    cycles in flush code, and bus bandwidth as all the lines of a
+	 *    buffer need to be flushed out to memory
+	 *   -For coherent data, Read/Write to buffers terminate early in cache
+	 *   (vs. always going to memory - thus are faster)
+	 */
+	if (is_isa_arcv2() && ioc_exists)
+		return dma_alloc_noncoherent(dev, size, dma_handle, gfp);
+
 	/* This is linear addr (0x8000_0000 based) */
 	paddr = alloc_pages_exact(size, gfp);
 	if (!paddr)
@@ -63,11 +75,23 @@ void *dma_alloc_coherent(struct device *dev, size_t size,
 
 	/* This is kernel Virtual address (0x7000_0000 based) */
 	kvaddr = ioremap_nocache((unsigned long)paddr, size);
-	if (kvaddr != NULL)
-		memset(kvaddr, 0, size);
+	if (kvaddr == NULL)
+		return NULL;
 
 	/* This is bus address, platform dependent */
-	*dma_handle = plat_kernel_addr_to_dma(dev, paddr);
+	*dma_handle = (dma_addr_t)paddr;
+
+	/*
+	 * Evict any existing L1 and/or L2 lines for the backing page
+	 * in case it was used earlier as a normal "cached" page.
+	 * Yeah this bit us - STAR 9000898266
+	 *
+	 * Although core does call flush_cache_vmap(), it gets kvaddr hence
+	 * can't be used to efficiently flush L1 and/or L2 which need paddr
+	 * Currently flush_cache_vmap nukes the L1 cache completely which
+	 * will be optimized as a separate commit
+	 */
+	dma_cache_wback_inv((unsigned long)paddr, size);
 
 	return kvaddr;
 }
@@ -76,10 +100,12 @@ EXPORT_SYMBOL(dma_alloc_coherent);
 void dma_free_coherent(struct device *dev, size_t size, void *kvaddr,
 		       dma_addr_t dma_handle)
 {
+	if (is_isa_arcv2() && ioc_exists)
+		return dma_free_noncoherent(dev, size, kvaddr, dma_handle);
+
 	iounmap((void __force __iomem *)kvaddr);
 
-	free_pages_exact((void *)plat_dma_addr_to_kernel(dev, dma_handle),
-			 size);
+	free_pages_exact((void *)dma_handle, size);
 }
 EXPORT_SYMBOL(dma_free_coherent);
 
diff --git a/kernel/arch/arc/mm/fault.c b/kernel/arch/arc/mm/fault.c
index d948e4e9d..af63f4a13 100644
--- a/kernel/arch/arc/mm/fault.c
+++ b/kernel/arch/arc/mm/fault.c
@@ -18,7 +18,14 @@
 #include <asm/pgalloc.h>
 #include <asm/mmu.h>
 
-static int handle_vmalloc_fault(unsigned long address)
+/*
+ * kernel virtual address is required to implement vmalloc/pkmap/fixmap
+ * Refer to asm/processor.h for System Memory Map
+ *
+ * It simply copies the PMD entry (pointer to 2nd level page table or hugepage)
+ * from swapper pgdir to task pgdir. The 2nd level table/page is thus shared
+ */
+noinline static int handle_kernel_vaddr_fault(unsigned long address)
 {
 	/*
 	 * Synchronize this task's top level page-table
@@ -72,8 +79,8 @@ void do_page_fault(unsigned long address, struct pt_regs *regs)
 	 * only copy the information from the master page table,
 	 * nothing more.
 	 */
-	if (address >= VMALLOC_START && address <= VMALLOC_END) {
-		ret = handle_vmalloc_fault(address);
+	if (address >= VMALLOC_START) {
+		ret = handle_kernel_vaddr_fault(address);
 		if (unlikely(ret))
 			goto bad_area_nosemaphore;
 		else
diff --git a/kernel/arch/arc/mm/highmem.c b/kernel/arch/arc/mm/highmem.c
new file mode 100644
index 000000000..92dd92cad
--- /dev/null
+++ b/kernel/arch/arc/mm/highmem.c
@@ -0,0 +1,140 @@
+/*
+ * Copyright (C) 2015 Synopsys, Inc. (www.synopsys.com)
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ */
+
+#include <linux/bootmem.h>
+#include <linux/export.h>
+#include <linux/highmem.h>
+#include <asm/processor.h>
+#include <asm/pgtable.h>
+#include <asm/pgalloc.h>
+#include <asm/tlbflush.h>
+
+/*
+ * HIGHMEM API:
+ *
+ * kmap() API provides sleep semantics hence refered to as "permanent maps"
+ * It allows mapping LAST_PKMAP pages, using @last_pkmap_nr as the cursor
+ * for book-keeping
+ *
+ * kmap_atomic() can't sleep (calls pagefault_disable()), thus it provides
+ * shortlived ala "temporary mappings" which historically were implemented as
+ * fixmaps (compile time addr etc). Their book-keeping is done per cpu.
+ *
+ *	Both these facts combined (preemption disabled and per-cpu allocation)
+ *	means the total number of concurrent fixmaps will be limited to max
+ *	such allocations in a single control path. Thus KM_TYPE_NR (another
+ *	historic relic) is a small'ish number which caps max percpu fixmaps
+ *
+ * ARC HIGHMEM Details
+ *
+ * - the kernel vaddr space from 0x7z to 0x8z (currently used by vmalloc/module)
+ *   is now shared between vmalloc and kmap (non overlapping though)
+ *
+ * - Both fixmap/pkmap use a dedicated page table each, hooked up to swapper PGD
+ *   This means each only has 1 PGDIR_SIZE worth of kvaddr mappings, which means
+ *   2M of kvaddr space for typical config (8K page and 11:8:13 traversal split)
+ *
+ * - fixmap anyhow needs a limited number of mappings. So 2M kvaddr == 256 PTE
+ *   slots across NR_CPUS would be more than sufficient (generic code defines
+ *   KM_TYPE_NR as 20).
+ *
+ * - pkmap being preemptible, in theory could do with more than 256 concurrent
+ *   mappings. However, generic pkmap code: map_new_virtual(), doesn't traverse
+ *   the PGD and only works with a single page table @pkmap_page_table, hence
+ *   sets the limit
+ */
+
+extern pte_t * pkmap_page_table;
+static pte_t * fixmap_page_table;
+
+void *kmap(struct page *page)
+{
+	BUG_ON(in_interrupt());
+	if (!PageHighMem(page))
+		return page_address(page);
+
+	return kmap_high(page);
+}
+
+void *kmap_atomic(struct page *page)
+{
+	int idx, cpu_idx;
+	unsigned long vaddr;
+
+	preempt_disable();
+	pagefault_disable();
+	if (!PageHighMem(page))
+		return page_address(page);
+
+	cpu_idx = kmap_atomic_idx_push();
+	idx = cpu_idx + KM_TYPE_NR * smp_processor_id();
+	vaddr = FIXMAP_ADDR(idx);
+
+	set_pte_at(&init_mm, vaddr, fixmap_page_table + idx,
+		   mk_pte(page, kmap_prot));
+
+	return (void *)vaddr;
+}
+EXPORT_SYMBOL(kmap_atomic);
+
+void __kunmap_atomic(void *kv)
+{
+	unsigned long kvaddr = (unsigned long)kv;
+
+	if (kvaddr >= FIXMAP_BASE && kvaddr < (FIXMAP_BASE + FIXMAP_SIZE)) {
+
+		/*
+		 * Because preemption is disabled, this vaddr can be associated
+		 * with the current allocated index.
+		 * But in case of multiple live kmap_atomic(), it still relies on
+		 * callers to unmap in right order.
+		 */
+		int cpu_idx = kmap_atomic_idx();
+		int idx = cpu_idx + KM_TYPE_NR * smp_processor_id();
+
+		WARN_ON(kvaddr != FIXMAP_ADDR(idx));
+
+		pte_clear(&init_mm, kvaddr, fixmap_page_table + idx);
+		local_flush_tlb_kernel_range(kvaddr, kvaddr + PAGE_SIZE);
+
+		kmap_atomic_idx_pop();
+	}
+
+	pagefault_enable();
+	preempt_enable();
+}
+EXPORT_SYMBOL(__kunmap_atomic);
+
+static noinline pte_t * __init alloc_kmap_pgtable(unsigned long kvaddr)
+{
+	pgd_t *pgd_k;
+	pud_t *pud_k;
+	pmd_t *pmd_k;
+	pte_t *pte_k;
+
+	pgd_k = pgd_offset_k(kvaddr);
+	pud_k = pud_offset(pgd_k, kvaddr);
+	pmd_k = pmd_offset(pud_k, kvaddr);
+
+	pte_k = (pte_t *)alloc_bootmem_low_pages(PAGE_SIZE);
+	pmd_populate_kernel(&init_mm, pmd_k, pte_k);
+	return pte_k;
+}
+
+void __init kmap_init(void)
+{
+	/* Due to recursive include hell, we can't do this in processor.h */
+	BUILD_BUG_ON(PAGE_OFFSET < (VMALLOC_END + FIXMAP_SIZE + PKMAP_SIZE));
+
+	BUILD_BUG_ON(KM_TYPE_NR > PTRS_PER_PTE);
+	pkmap_page_table = alloc_kmap_pgtable(PKMAP_BASE);
+
+	BUILD_BUG_ON(LAST_PKMAP > PTRS_PER_PTE);
+	fixmap_page_table = alloc_kmap_pgtable(FIXMAP_BASE);
+}
diff --git a/kernel/arch/arc/mm/init.c b/kernel/arch/arc/mm/init.c
index d44eedd8c..7d2c4fbf4 100644
--- a/kernel/arch/arc/mm/init.c
+++ b/kernel/arch/arc/mm/init.c
@@ -15,6 +15,7 @@
 #endif
 #include <linux/swap.h>
 #include <linux/module.h>
+#include <linux/highmem.h>
 #include <asm/page.h>
 #include <asm/pgalloc.h>
 #include <asm/sections.h>
@@ -24,16 +25,22 @@ pgd_t swapper_pg_dir[PTRS_PER_PGD] __aligned(PAGE_SIZE);
 char empty_zero_page[PAGE_SIZE] __aligned(PAGE_SIZE);
 EXPORT_SYMBOL(empty_zero_page);
 
-/* Default tot mem from .config */
-static unsigned long arc_mem_sz = 0x20000000;  /* some default */
+static const unsigned long low_mem_start = CONFIG_LINUX_LINK_BASE;
+static unsigned long low_mem_sz;
+
+#ifdef CONFIG_HIGHMEM
+static unsigned long min_high_pfn;
+static u64 high_mem_start;
+static u64 high_mem_sz;
+#endif
 
 /* User can over-ride above with "mem=nnn[KkMm]" in cmdline */
 static int __init setup_mem_sz(char *str)
 {
-	arc_mem_sz = memparse(str, NULL) & PAGE_MASK;
+	low_mem_sz = memparse(str, NULL) & PAGE_MASK;
 
 	/* early console might not be setup yet - it will show up later */
-	pr_info("\"mem=%s\": mem sz set to %ldM\n", str, TO_MB(arc_mem_sz));
+	pr_info("\"mem=%s\": mem sz set to %ldM\n", str, TO_MB(low_mem_sz));
 
 	return 0;
 }
@@ -41,8 +48,24 @@ early_param("mem", setup_mem_sz);
 
 void __init early_init_dt_add_memory_arch(u64 base, u64 size)
 {
-	arc_mem_sz = size & PAGE_MASK;
-	pr_info("Memory size set via devicetree %ldM\n", TO_MB(arc_mem_sz));
+	int in_use = 0;
+
+	if (!low_mem_sz) {
+		if (base != low_mem_start)
+			panic("CONFIG_LINUX_LINK_BASE != DT memory { }");
+
+		low_mem_sz = size;
+		in_use = 1;
+	} else {
+#ifdef CONFIG_HIGHMEM
+		high_mem_start = base;
+		high_mem_sz = size;
+		in_use = 1;
+#endif
+	}
+
+	pr_info("Memory @ %llx [%lldM] %s\n",
+		base, TO_MB(size), !in_use ? "Not used":"");
 }
 
 #ifdef CONFIG_BLK_DEV_INITRD
@@ -72,46 +95,62 @@ early_param("initrd", early_initrd);
 void __init setup_arch_memory(void)
 {
 	unsigned long zones_size[MAX_NR_ZONES];
-	unsigned long end_mem = CONFIG_LINUX_LINK_BASE + arc_mem_sz;
+	unsigned long zones_holes[MAX_NR_ZONES];
 
 	init_mm.start_code = (unsigned long)_text;
 	init_mm.end_code = (unsigned long)_etext;
 	init_mm.end_data = (unsigned long)_edata;
 	init_mm.brk = (unsigned long)_end;
 
-	/*
-	 * We do it here, so that memory is correctly instantiated
-	 * even if "mem=xxx" cmline over-ride is given and/or
-	 * DT has memory node. Each causes an update to @arc_mem_sz
-	 * and we finally add memory one here
-	 */
-	memblock_add(CONFIG_LINUX_LINK_BASE, arc_mem_sz);
-
-	/*------------- externs in mm need setting up ---------------*/
-
 	/* first page of system - kernel .vector starts here */
 	min_low_pfn = ARCH_PFN_OFFSET;
 
-	/* Last usable page of low mem (no HIGHMEM yet for ARC port) */
-	max_low_pfn = max_pfn = PFN_DOWN(end_mem);
+	/* Last usable page of low mem */
+	max_low_pfn = max_pfn = PFN_DOWN(low_mem_start + low_mem_sz);
 
-	max_mapnr = max_low_pfn - min_low_pfn;
+#ifdef CONFIG_HIGHMEM
+	min_high_pfn = PFN_DOWN(high_mem_start);
+	max_pfn = PFN_DOWN(high_mem_start + high_mem_sz);
+#endif
+
+	max_mapnr = max_pfn - min_low_pfn;
 
-	/*------------- reserve kernel image -----------------------*/
-	memblock_reserve(CONFIG_LINUX_LINK_BASE,
-			 __pa(_end) - CONFIG_LINUX_LINK_BASE);
+	/*------------- bootmem allocator setup -----------------------*/
+
+	/*
+	 * seed the bootmem allocator after any DT memory node parsing or
+	 * "mem=xxx" cmdline overrides have potentially updated @arc_mem_sz
+	 *
+	 * Only low mem is added, otherwise we have crashes when allocating
+	 * mem_map[] itself. NO_BOOTMEM allocates mem_map[] at the end of
+	 * avail memory, ending in highmem with a > 32-bit address. However
+	 * it then tries to memset it with a truncaed 32-bit handle, causing
+	 * the crash
+	 */
+
+	memblock_add(low_mem_start, low_mem_sz);
+	memblock_reserve(low_mem_start, __pa(_end) - low_mem_start);
 
 #ifdef CONFIG_BLK_DEV_INITRD
-	/*------------- reserve initrd image -----------------------*/
 	if (initrd_start)
 		memblock_reserve(__pa(initrd_start), initrd_end - initrd_start);
 #endif
 
 	memblock_dump_all();
 
-	/*-------------- node setup --------------------------------*/
+	/*----------------- node/zones setup --------------------------*/
 	memset(zones_size, 0, sizeof(zones_size));
-	zones_size[ZONE_NORMAL] = max_mapnr;
+	memset(zones_holes, 0, sizeof(zones_holes));
+
+	zones_size[ZONE_NORMAL] = max_low_pfn - min_low_pfn;
+	zones_holes[ZONE_NORMAL] = 0;
+
+#ifdef CONFIG_HIGHMEM
+	zones_size[ZONE_HIGHMEM] = max_pfn - max_low_pfn;
+
+	/* This handles the peripheral address space hole */
+	zones_holes[ZONE_HIGHMEM] = min_high_pfn - max_low_pfn;
+#endif
 
 	/*
 	 * We can't use the helper free_area_init(zones[]) because it uses
@@ -122,9 +161,12 @@ void __init setup_arch_memory(void)
 	free_area_init_node(0,			/* node-id */
 			    zones_size,		/* num pages per zone */
 			    min_low_pfn,	/* first pfn of node */
-			    NULL);		/* NO holes */
+			    zones_holes);	/* holes */
 
-	high_memory = (void *)end_mem;
+#ifdef CONFIG_HIGHMEM
+	high_memory = (void *)(min_high_pfn << PAGE_SHIFT);
+	kmap_init();
+#endif
 }
 
 /*
@@ -135,6 +177,14 @@ void __init setup_arch_memory(void)
  */
 void __init mem_init(void)
 {
+#ifdef CONFIG_HIGHMEM
+	unsigned long tmp;
+
+	reset_all_zones_managed_pages();
+	for (tmp = min_high_pfn; tmp < max_pfn; tmp++)
+		free_highmem_page(pfn_to_page(tmp));
+#endif
+
 	free_all_bootmem();
 	mem_init_print_info(NULL);
 }
diff --git a/kernel/arch/arc/mm/tlb.c b/kernel/arch/arc/mm/tlb.c
index 7f47d2a56..daf2bf52b 100644
--- a/kernel/arch/arc/mm/tlb.c
+++ b/kernel/arch/arc/mm/tlb.c
@@ -109,10 +109,16 @@ DEFINE_PER_CPU(unsigned int, asid_cache) = MM_CTXT_FIRST_CYCLE;
 static inline void __tlb_entry_erase(void)
 {
 	write_aux_reg(ARC_REG_TLBPD1, 0);
+
+	if (is_pae40_enabled())
+		write_aux_reg(ARC_REG_TLBPD1HI, 0);
+
 	write_aux_reg(ARC_REG_TLBPD0, 0);
 	write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
 }
 
+#if (CONFIG_ARC_MMU_VER < 4)
+
 static inline unsigned int tlb_entry_lkup(unsigned long vaddr_n_asid)
 {
 	unsigned int idx;
@@ -180,7 +186,7 @@ static void utlb_invalidate(void)
 
 }
 
-static void tlb_entry_insert(unsigned int pd0, unsigned int pd1)
+static void tlb_entry_insert(unsigned int pd0, pte_t pd1)
 {
 	unsigned int idx;
 
@@ -210,28 +216,71 @@ static void tlb_entry_insert(unsigned int pd0, unsigned int pd1)
 	write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
 }
 
+#else	/* CONFIG_ARC_MMU_VER >= 4) */
+
+static void utlb_invalidate(void)
+{
+	/* No need since uTLB is always in sync with JTLB */
+}
+
+static void tlb_entry_erase(unsigned int vaddr_n_asid)
+{
+	write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid | _PAGE_PRESENT);
+	write_aux_reg(ARC_REG_TLBCOMMAND, TLBDeleteEntry);
+}
+
+static void tlb_entry_insert(unsigned int pd0, pte_t pd1)
+{
+	write_aux_reg(ARC_REG_TLBPD0, pd0);
+	write_aux_reg(ARC_REG_TLBPD1, pd1);
+
+	if (is_pae40_enabled())
+		write_aux_reg(ARC_REG_TLBPD1HI, (u64)pd1 >> 32);
+
+	write_aux_reg(ARC_REG_TLBCOMMAND, TLBInsertEntry);
+}
+
+#endif
+
 /*
  * Un-conditionally (without lookup) erase the entire MMU contents
  */
 
 noinline void local_flush_tlb_all(void)
 {
+	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
 	unsigned long flags;
 	unsigned int entry;
-	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
+	int num_tlb = mmu->sets * mmu->ways;
 
 	local_irq_save(flags);
 
 	/* Load PD0 and PD1 with template for a Blank Entry */
 	write_aux_reg(ARC_REG_TLBPD1, 0);
+
+	if (is_pae40_enabled())
+		write_aux_reg(ARC_REG_TLBPD1HI, 0);
+
 	write_aux_reg(ARC_REG_TLBPD0, 0);
 
-	for (entry = 0; entry < mmu->num_tlb; entry++) {
+	for (entry = 0; entry < num_tlb; entry++) {
 		/* write this entry to the TLB */
 		write_aux_reg(ARC_REG_TLBINDEX, entry);
 		write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
 	}
 
+	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
+		const int stlb_idx = 0x800;
+
+		/* Blank sTLB entry */
+		write_aux_reg(ARC_REG_TLBPD0, _PAGE_HW_SZ);
+
+		for (entry = stlb_idx; entry < stlb_idx + 16; entry++) {
+			write_aux_reg(ARC_REG_TLBINDEX, entry);
+			write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
+		}
+	}
+
 	utlb_invalidate();
 
 	local_irq_restore(flags);
@@ -385,6 +434,15 @@ static inline void ipi_flush_tlb_range(void *arg)
 	local_flush_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
 }
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+static inline void ipi_flush_pmd_tlb_range(void *arg)
+{
+	struct tlb_args *ta = arg;
+
+	local_flush_pmd_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
+}
+#endif
+
 static inline void ipi_flush_tlb_kernel_range(void *arg)
 {
 	struct tlb_args *ta = (struct tlb_args *)arg;
@@ -425,6 +483,20 @@ void flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
 	on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_range, &ta, 1);
 }
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+void flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
+			 unsigned long end)
+{
+	struct tlb_args ta = {
+		.ta_vma = vma,
+		.ta_start = start,
+		.ta_end = end
+	};
+
+	on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_pmd_tlb_range, &ta, 1);
+}
+#endif
+
 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 {
 	struct tlb_args ta = {
@@ -439,11 +511,12 @@ void flush_tlb_kernel_range(unsigned long start, unsigned long end)
 /*
  * Routine to create a TLB entry
  */
-void create_tlb(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
+void create_tlb(struct vm_area_struct *vma, unsigned long vaddr, pte_t *ptep)
 {
 	unsigned long flags;
 	unsigned int asid_or_sasid, rwx;
-	unsigned long pd0, pd1;
+	unsigned long pd0;
+	pte_t pd1;
 
 	/*
 	 * create_tlb() assumes that current->mm == vma->mm, since
@@ -475,9 +548,9 @@ void create_tlb(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
 
 	local_irq_save(flags);
 
-	tlb_paranoid_check(asid_mm(vma->vm_mm, smp_processor_id()), address);
+	tlb_paranoid_check(asid_mm(vma->vm_mm, smp_processor_id()), vaddr);
 
-	address &= PAGE_MASK;
+	vaddr &= PAGE_MASK;
 
 	/* update this PTE credentials */
 	pte_val(*ptep) |= (_PAGE_PRESENT | _PAGE_ACCESSED);
@@ -487,7 +560,7 @@ void create_tlb(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
 	/* ASID for this task */
 	asid_or_sasid = read_aux_reg(ARC_REG_PID) & 0xff;
 
-	pd0 = address | asid_or_sasid | (pte_val(*ptep) & PTE_BITS_IN_PD0);
+	pd0 = vaddr | asid_or_sasid | (pte_val(*ptep) & PTE_BITS_IN_PD0);
 
 	/*
 	 * ARC MMU provides fully orthogonal access bits for K/U mode,
@@ -523,7 +596,7 @@ void update_mmu_cache(struct vm_area_struct *vma, unsigned long vaddr_unaligned,
 		      pte_t *ptep)
 {
 	unsigned long vaddr = vaddr_unaligned & PAGE_MASK;
-	unsigned long paddr = pte_val(*ptep) & PAGE_MASK;
+	phys_addr_t paddr = pte_val(*ptep) & PAGE_MASK;
 	struct page *page = pfn_to_page(pte_pfn(*ptep));
 
 	create_tlb(vma, vaddr, ptep);
@@ -546,16 +619,105 @@ void update_mmu_cache(struct vm_area_struct *vma, unsigned long vaddr_unaligned,
 
 		int dirty = !test_and_set_bit(PG_dc_clean, &page->flags);
 		if (dirty) {
-			/* wback + inv dcache lines */
+			/* wback + inv dcache lines (K-mapping) */
 			__flush_dcache_page(paddr, paddr);
 
-			/* invalidate any existing icache lines */
+			/* invalidate any existing icache lines (U-mapping) */
 			if (vma->vm_flags & VM_EXEC)
 				__inv_icache_page(paddr, vaddr);
 		}
 	}
 }
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+
+/*
+ * MMUv4 in HS38x cores supports Super Pages which are basis for Linux THP
+ * support.
+ *
+ * Normal and Super pages can co-exist (ofcourse not overlap) in TLB with a
+ * new bit "SZ" in TLB page desciptor to distinguish between them.
+ * Super Page size is configurable in hardware (4K to 16M), but fixed once
+ * RTL builds.
+ *
+ * The exact THP size a Linx configuration will support is a function of:
+ *  - MMU page size (typical 8K, RTL fixed)
+ *  - software page walker address split between PGD:PTE:PFN (typical
+ *    11:8:13, but can be changed with 1 line)
+ * So for above default, THP size supported is 8K * (2^8) = 2M
+ *
+ * Default Page Walker is 2 levels, PGD:PTE:PFN, which in THP regime
+ * reduces to 1 level (as PTE is folded into PGD and canonically referred
+ * to as PMD).
+ * Thus THP PMD accessors are implemented in terms of PTE (just like sparc)
+ */
+
+void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
+				 pmd_t *pmd)
+{
+	pte_t pte = __pte(pmd_val(*pmd));
+	update_mmu_cache(vma, addr, &pte);
+}
+
+void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
+				pgtable_t pgtable)
+{
+	struct list_head *lh = (struct list_head *) pgtable;
+
+	assert_spin_locked(&mm->page_table_lock);
+
+	/* FIFO */
+	if (!pmd_huge_pte(mm, pmdp))
+		INIT_LIST_HEAD(lh);
+	else
+		list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
+	pmd_huge_pte(mm, pmdp) = pgtable;
+}
+
+pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
+{
+	struct list_head *lh;
+	pgtable_t pgtable;
+
+	assert_spin_locked(&mm->page_table_lock);
+
+	pgtable = pmd_huge_pte(mm, pmdp);
+	lh = (struct list_head *) pgtable;
+	if (list_empty(lh))
+		pmd_huge_pte(mm, pmdp) = NULL;
+	else {
+		pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
+		list_del(lh);
+	}
+
+	pte_val(pgtable[0]) = 0;
+	pte_val(pgtable[1]) = 0;
+
+	return pgtable;
+}
+
+void local_flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
+			       unsigned long end)
+{
+	unsigned int cpu;
+	unsigned long flags;
+
+	local_irq_save(flags);
+
+	cpu = smp_processor_id();
+
+	if (likely(asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID)) {
+		unsigned int asid = hw_pid(vma->vm_mm, cpu);
+
+		/* No need to loop here: this will always be for 1 Huge Page */
+		tlb_entry_erase(start | _PAGE_HW_SZ | asid);
+	}
+
+	local_irq_restore(flags);
+}
+
+#endif
+
 /* Read the Cache Build Confuration Registers, Decode them and save into
  * the cpuinfo structure for later use.
  * No Validation is done here, simply read/convert the BCRs
@@ -574,47 +736,73 @@ void read_decode_mmu_bcr(void)
 
 	struct bcr_mmu_3 {
 #ifdef CONFIG_CPU_BIG_ENDIAN
-	unsigned int ver:8, ways:4, sets:4, osm:1, reserv:3, pg_sz:4,
+	unsigned int ver:8, ways:4, sets:4, res:3, sasid:1, pg_sz:4,
 		     u_itlb:4, u_dtlb:4;
 #else
-	unsigned int u_dtlb:4, u_itlb:4, pg_sz:4, reserv:3, osm:1, sets:4,
+	unsigned int u_dtlb:4, u_itlb:4, pg_sz:4, sasid:1, res:3, sets:4,
 		     ways:4, ver:8;
 #endif
 	} *mmu3;
 
+	struct bcr_mmu_4 {
+#ifdef CONFIG_CPU_BIG_ENDIAN
+	unsigned int ver:8, sasid:1, sz1:4, sz0:4, res:2, pae:1,
+		     n_ways:2, n_entry:2, n_super:2, u_itlb:3, u_dtlb:3;
+#else
+	/*           DTLB      ITLB      JES        JE         JA      */
+	unsigned int u_dtlb:3, u_itlb:3, n_super:2, n_entry:2, n_ways:2,
+		     pae:1, res:2, sz0:4, sz1:4, sasid:1, ver:8;
+#endif
+	} *mmu4;
+
 	tmp = read_aux_reg(ARC_REG_MMU_BCR);
 	mmu->ver = (tmp >> 24);
 
 	if (mmu->ver <= 2) {
 		mmu2 = (struct bcr_mmu_1_2 *)&tmp;
-		mmu->pg_sz = PAGE_SIZE;
+		mmu->pg_sz_k = TO_KB(0x2000);
 		mmu->sets = 1 << mmu2->sets;
 		mmu->ways = 1 << mmu2->ways;
 		mmu->u_dtlb = mmu2->u_dtlb;
 		mmu->u_itlb = mmu2->u_itlb;
-	} else {
+	} else if (mmu->ver == 3) {
 		mmu3 = (struct bcr_mmu_3 *)&tmp;
-		mmu->pg_sz = 512 << mmu3->pg_sz;
+		mmu->pg_sz_k = 1 << (mmu3->pg_sz - 1);
 		mmu->sets = 1 << mmu3->sets;
 		mmu->ways = 1 << mmu3->ways;
 		mmu->u_dtlb = mmu3->u_dtlb;
 		mmu->u_itlb = mmu3->u_itlb;
+		mmu->sasid = mmu3->sasid;
+	} else {
+		mmu4 = (struct bcr_mmu_4 *)&tmp;
+		mmu->pg_sz_k = 1 << (mmu4->sz0 - 1);
+		mmu->s_pg_sz_m = 1 << (mmu4->sz1 - 11);
+		mmu->sets = 64 << mmu4->n_entry;
+		mmu->ways = mmu4->n_ways * 2;
+		mmu->u_dtlb = mmu4->u_dtlb * 4;
+		mmu->u_itlb = mmu4->u_itlb * 4;
+		mmu->sasid = mmu4->sasid;
+		mmu->pae = mmu4->pae;
 	}
-
-	mmu->num_tlb = mmu->sets * mmu->ways;
 }
 
 char *arc_mmu_mumbojumbo(int cpu_id, char *buf, int len)
 {
 	int n = 0;
 	struct cpuinfo_arc_mmu *p_mmu = &cpuinfo_arc700[cpu_id].mmu;
+	char super_pg[64] = "";
+
+	if (p_mmu->s_pg_sz_m)
+		scnprintf(super_pg, 64, "%dM Super Page%s, ",
+			  p_mmu->s_pg_sz_m,
+			  IS_USED_CFG(CONFIG_TRANSPARENT_HUGEPAGE));
 
 	n += scnprintf(buf + n, len - n,
-		      "MMU [v%x]\t: %dk PAGE, JTLB %d (%dx%d), uDTLB %d, uITLB %d %s\n",
-		       p_mmu->ver, TO_KB(p_mmu->pg_sz),
-		       p_mmu->num_tlb, p_mmu->sets, p_mmu->ways,
+		      "MMU [v%x]\t: %dk PAGE, %sJTLB %d (%dx%d), uDTLB %d, uITLB %d %s%s\n",
+		       p_mmu->ver, p_mmu->pg_sz_k, super_pg,
+		       p_mmu->sets * p_mmu->ways, p_mmu->sets, p_mmu->ways,
 		       p_mmu->u_dtlb, p_mmu->u_itlb,
-		       IS_ENABLED(CONFIG_ARC_MMU_SASID) ? ",SASID" : "");
+		       IS_AVAIL2(p_mmu->pae, "PAE40 ", CONFIG_ARC_HAS_PAE40));
 
 	return buf;
 }
@@ -639,9 +827,17 @@ void arc_mmu_init(void)
 		      mmu->ver, CONFIG_ARC_MMU_VER);
 	}
 
-	if (mmu->pg_sz != PAGE_SIZE)
+	if (mmu->pg_sz_k != TO_KB(PAGE_SIZE))
 		panic("MMU pg size != PAGE_SIZE (%luk)\n", TO_KB(PAGE_SIZE));
 
+	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
+	    mmu->s_pg_sz_m != TO_MB(HPAGE_PMD_SIZE))
+		panic("MMU Super pg size != Linux HPAGE_PMD_SIZE (%luM)\n",
+		      (unsigned long)TO_MB(HPAGE_PMD_SIZE));
+
+	if (IS_ENABLED(CONFIG_ARC_HAS_PAE40) && !mmu->pae)
+		panic("Hardware doesn't support PAE40\n");
+
 	/* Enable the MMU */
 	write_aux_reg(ARC_REG_PID, MMU_ENABLE);
 
@@ -677,15 +873,15 @@ void arc_mmu_init(void)
  *      the duplicate one.
  * -Knob to be verbose abt it.(TODO: hook them up to debugfs)
  */
-volatile int dup_pd_verbose = 1;/* Be slient abt it or complain (default) */
+volatile int dup_pd_silent; /* Be slient abt it or complain (default) */
 
 void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
 			  struct pt_regs *regs)
 {
-	int set, way, n;
-	unsigned long flags, is_valid;
 	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
-	unsigned int pd0[mmu->ways], pd1[mmu->ways];
+	unsigned int pd0[mmu->ways];
+	unsigned long flags;
+	int set;
 
 	local_irq_save(flags);
 
@@ -695,14 +891,16 @@ void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
 	/* loop thru all sets of TLB */
 	for (set = 0; set < mmu->sets; set++) {
 
+		int is_valid, way;
+
 		/* read out all the ways of current set */
 		for (way = 0, is_valid = 0; way < mmu->ways; way++) {
 			write_aux_reg(ARC_REG_TLBINDEX,
 					  SET_WAY_TO_IDX(mmu, set, way));
 			write_aux_reg(ARC_REG_TLBCOMMAND, TLBRead);
 			pd0[way] = read_aux_reg(ARC_REG_TLBPD0);
-			pd1[way] = read_aux_reg(ARC_REG_TLBPD1);
 			is_valid |= pd0[way] & _PAGE_PRESENT;
+			pd0[way] &= PAGE_MASK;
 		}
 
 		/* If all the WAYS in SET are empty, skip to next SET */
@@ -711,30 +909,28 @@ void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
 
 		/* Scan the set for duplicate ways: needs a nested loop */
 		for (way = 0; way < mmu->ways - 1; way++) {
+
+			int n;
+
 			if (!pd0[way])
 				continue;
 
 			for (n = way + 1; n < mmu->ways; n++) {
-				if ((pd0[way] & PAGE_MASK) ==
-				    (pd0[n] & PAGE_MASK)) {
-
-					if (dup_pd_verbose) {
-						pr_info("Duplicate PD's @"
-							"[%d:%d]/[%d:%d]\n",
-						     set, way, set, n);
-						pr_info("TLBPD0[%u]: %08x\n",
-						     way, pd0[way]);
-					}
-
-					/*
-					 * clear entry @way and not @n. This is
-					 * critical to our optimised loop
-					 */
-					pd0[way] = pd1[way] = 0;
-					write_aux_reg(ARC_REG_TLBINDEX,
+				if (pd0[way] != pd0[n])
+					continue;
+
+				if (!dup_pd_silent)
+					pr_info("Dup TLB PD0 %08x @ set %d ways %d,%d\n",
+						pd0[way], set, way, n);
+
+				/*
+				 * clear entry @way and not @n.
+				 * This is critical to our optimised loop
+				 */
+				pd0[way] = 0;
+				write_aux_reg(ARC_REG_TLBINDEX,
 						SET_WAY_TO_IDX(mmu, set, way));
-					__tlb_entry_erase();
-				}
+				__tlb_entry_erase();
 			}
 		}
 	}
diff --git a/kernel/arch/arc/mm/tlbex.S b/kernel/arch/arc/mm/tlbex.S
index d572f1c2c..f1967eeb3 100644
--- a/kernel/arch/arc/mm/tlbex.S
+++ b/kernel/arch/arc/mm/tlbex.S
@@ -35,8 +35,6 @@
  * Rahul Trivedi, Amit Bhor: Codito Technologies 2004
  */
 
-	.cpu A7
-
 #include <linux/linkage.h>
 #include <asm/entry.h>
 #include <asm/mmu.h>
@@ -46,6 +44,7 @@
 #include <asm/processor.h>
 #include <asm/tlb-mmu1.h>
 
+#ifdef CONFIG_ISA_ARCOMPACT
 ;-----------------------------------------------------------------
 ; ARC700 Exception Handling doesn't auto-switch stack and it only provides
 ; ONE scratch AUX reg "ARC_REG_SCRATCH_DATA0"
@@ -89,7 +88,7 @@ ex_saved_reg1:
 #ifdef CONFIG_SMP
 	sr  r0, [ARC_REG_SCRATCH_DATA0]	; freeup r0 to code with
 	GET_CPU_ID  r0			; get to per cpu scratch mem,
-	lsl r0, r0, L1_CACHE_SHIFT	; cache line wide per cpu
+	asl r0, r0, L1_CACHE_SHIFT	; cache line wide per cpu
 	add r0, @ex_saved_reg1, r0
 #else
 	st    r0, [@ex_saved_reg1]
@@ -108,7 +107,7 @@ ex_saved_reg1:
 .macro TLBMISS_RESTORE_REGS
 #ifdef CONFIG_SMP
 	GET_CPU_ID  r0			; get to per cpu scratch mem
-	lsl r0, r0, L1_CACHE_SHIFT	; each is cache line wide
+	asl r0, r0, L1_CACHE_SHIFT	; each is cache line wide
 	add r0, @ex_saved_reg1, r0
 	ld_s  r3, [r0,12]
 	ld_s  r2, [r0, 8]
@@ -123,6 +122,24 @@ ex_saved_reg1:
 #endif
 .endm
 
+#else	/* ARCv2 */
+
+.macro TLBMISS_FREEUP_REGS
+	PUSH  r0
+	PUSH  r1
+	PUSH  r2
+	PUSH  r3
+.endm
+
+.macro TLBMISS_RESTORE_REGS
+	POP   r3
+	POP   r2
+	POP   r1
+	POP   r0
+.endm
+
+#endif
+
 ;============================================================================
 ;  Troubleshooting Stuff
 ;============================================================================
@@ -188,20 +205,38 @@ ex_saved_reg1:
 #endif
 
 	lsr     r0, r2, PGDIR_SHIFT     ; Bits for indexing into PGD
-	ld.as   r1, [r1, r0]            ; PGD entry corresp to faulting addr
-	and.f   r1, r1, PAGE_MASK       ; Ignoring protection and other flags
-	;   contains Ptr to Page Table
-	bz.d    do_slow_path_pf         ; if no Page Table, do page fault
+	ld.as   r3, [r1, r0]            ; PGD entry corresp to faulting addr
+	tst	r3, r3
+	bz	do_slow_path_pf         ; if no Page Table, do page fault
+
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	and.f	0, r3, _PAGE_HW_SZ	; Is this Huge PMD (thp)
+	add2.nz	r1, r1, r0
+	bnz.d	2f		; YES: PGD == PMD has THP PTE: stop pgd walk
+	mov.nz	r0, r3
+
+#endif
+	and	r1, r3, PAGE_MASK
 
 	; Get the PTE entry: The idea is
 	; (1) x = addr >> PAGE_SHIFT 	-> masks page-off bits from @fault-addr
 	; (2) y = x & (PTRS_PER_PTE - 1) -> to get index
-	; (3) z = pgtbl[y]
-	; To avoid the multiply by in end, we do the -2, <<2 below
+	; (3) z = (pgtbl + y * 4)
+
+#ifdef CONFIG_ARC_HAS_PAE40
+#define PTE_SIZE_LOG	3	/* 8 == 2 ^ 3 */
+#else
+#define PTE_SIZE_LOG	2	/* 4 == 2 ^ 2 */
+#endif
+
+	; multiply in step (3) above avoided by shifting lesser in step (1)
+	lsr     r0, r2, ( PAGE_SHIFT - PTE_SIZE_LOG )
+	and     r0, r0, ( (PTRS_PER_PTE - 1) << PTE_SIZE_LOG )
+	ld.aw   r0, [r1, r0]            ; r0: PTE (lower word only for PAE40)
+					; r1: PTE ptr
+
+2:
 
-	lsr     r0, r2, (PAGE_SHIFT - 2)
-	and     r0, r0, ( (PTRS_PER_PTE - 1) << 2)
-	ld.aw   r0, [r1, r0]            ; get PTE and PTE ptr for fault addr
 #ifdef CONFIG_ARC_DBG_TLB_MISS_COUNT
 	and.f 0, r0, _PAGE_PRESENT
 	bz   1f
@@ -216,18 +251,23 @@ ex_saved_reg1:
 ;-----------------------------------------------------------------
 ; Convert Linux PTE entry into TLB entry
 ; A one-word PTE entry is programmed as two-word TLB Entry [PD0:PD1] in mmu
+;    (for PAE40, two-words PTE, while three-word TLB Entry [PD0:PD1:PD1HI])
 ; IN: r0 = PTE, r1 = ptr to PTE
 
 .macro CONV_PTE_TO_TLB
-	and    r3, r0, PTE_BITS_RWX	;       r w x
-	lsl    r2, r3, 3		; r w x 0 0 0 (GLOBAL, kernel only)
+	and    r3, r0, PTE_BITS_RWX	;          r  w  x
+	asl    r2, r3, 3		; Kr Kw Kx 0  0  0 (GLOBAL, kernel only)
 	and.f  0,  r0, _PAGE_GLOBAL
-	or.z   r2, r2, r3		; r w x r w x (!GLOBAL, user page)
+	or.z   r2, r2, r3		; Kr Kw Kx Ur Uw Ux (!GLOBAL, user page)
 
 	and r3, r0, PTE_BITS_NON_RWX_IN_PD1 ; Extract PFN+cache bits from PTE
 	or  r3, r3, r2
 
-	sr  r3, [ARC_REG_TLBPD1]    	; these go in PD1
+	sr  r3, [ARC_REG_TLBPD1]    	; paddr[31..13] | Kr Kw Kx Ur Uw Ux | C
+#ifdef	CONFIG_ARC_HAS_PAE40
+	ld	r3, [r1, 4]		; paddr[39..32]
+	sr	r3, [ARC_REG_TLBPD1HI]
+#endif
 
 	and r2, r0, PTE_BITS_IN_PD0 ; Extract other PTE flags: (V)alid, (G)lb
 
@@ -241,6 +281,7 @@ ex_saved_reg1:
 ; Commit the TLB entry into MMU
 
 .macro COMMIT_ENTRY_TO_MMU
+#if (CONFIG_ARC_MMU_VER < 4)
 
 	/* Get free TLB slot: Set = computed from vaddr, way = random */
 	sr  TLBGetIndex, [ARC_REG_TLBCOMMAND]
@@ -251,6 +292,10 @@ ex_saved_reg1:
 #else
 	sr TLBWrite, [ARC_REG_TLBCOMMAND]
 #endif
+
+#else
+	sr TLBInsertEntry, [ARC_REG_TLBCOMMAND]
+#endif
 .endm
 
 
@@ -291,6 +336,7 @@ ENTRY(EV_TLBMissI)
 	CONV_PTE_TO_TLB
 	COMMIT_ENTRY_TO_MMU
 	TLBMISS_RESTORE_REGS
+EV_TLBMissI_fast_ret:	; additional label for VDK OS-kit instrumentation
 	rtie
 
 END(EV_TLBMissI)
@@ -342,7 +388,7 @@ ENTRY(EV_TLBMissD)
 	lr      r3, [ecr]
 	or      r0, r0, _PAGE_ACCESSED        ; Accessed bit always
 	btst_s  r3,  ECR_C_BIT_DTLB_ST_MISS   ; See if it was a Write Access ?
-	or.nz   r0, r0, _PAGE_MODIFIED        ; if Write, set Dirty bit as well
+	or.nz   r0, r0, _PAGE_DIRTY           ; if Write, set Dirty bit as well
 	st_s    r0, [r1]                      ; Write back PTE
 
 	CONV_PTE_TO_TLB
@@ -356,6 +402,7 @@ ENTRY(EV_TLBMissD)
 
 	COMMIT_ENTRY_TO_MMU
 	TLBMISS_RESTORE_REGS
+EV_TLBMissD_fast_ret:	; additional label for VDK OS-kit instrumentation
 	rtie
 
 ;-------- Common routine to call Linux Page Fault Handler -----------
@@ -366,19 +413,5 @@ do_slow_path_pf:
 
 	; Slow path TLB Miss handled as a regular ARC Exception
 	; (stack switching / save the complete reg-file).
-	EXCEPTION_PROLOGUE
-
-	; ------- setup args for Linux Page fault Hanlder ---------
-	mov_s r1, sp
-	lr    r0, [efa]
-
-	; We don't want exceptions to be disabled while the fault is handled.
-	; Now that we have saved the context we return from exception hence
-	; exceptions get re-enable
-
-	FAKE_RET_FROM_EXCPN  r9
-
-	bl  do_page_fault
-	b   ret_from_exception
-
+	b  call_do_page_fault
 END(EV_TLBMissD)